Electromagnetic modeling of PMSM based on multi-scale physics-informed neural network with dual-level Bayesian uncertainty weighting mechanism

Accurate estimation of electromagnetic responses in permanent magnet synchronous motors (PMSMs) is crucial for complex engineering applications. Therefore, a multi-scale physics-informed neural network (M-BPINN) with a dual-level Bayesian uncertainty weighting mechanism is proposed. Discrete spatial scaling and periodic sinusoidal representation networks (SIRENs) are utilized by the multi-scale architecture to ensure that high-frequency spatial features are accurately captured. The adaptive weighting mechanism based on Bayesian uncertainty will automatically evaluate the reliability of constraints to reduce gradient stiffness. The evaluation was conducted on the simulation model of an 8-pole 54 slot conical rotor permanent magnet synchronous motor, the relative average absolute error (RelMAE) of magnetic vector potential A_z is less than 2.5%, and the RelMAE of magnetic flux density B and field strength H is 7%-13%, demonstrating improved accuracy compared to traditional methods. Furthermore, computational speed is accelerated by three orders of magnitude by this model. It is demonstrated that significant potential for real-time control and digital twin applications.